Research map / Continual learning / Online and replay-based
DualNet: Continual Learning, Fast and Slow
Quang Pham, Chenghao Liu, Steven HOI
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DualNet's separate slow self-supervised learner and fast adaptation learner | Continual learning with reduced ResNet18 · Split miniImageNet-TA | ACC: 68.04±1.23 -> 73.20±0.68 | Improved |
| Use DualNet's separate slow self-supervised learner and fast adaptation learner | Continual learning with reduced ResNet18 · Split miniImageNet-TF | ACC: 25.26±1.81 -> 37.56±0.52 | Improved |
| Use DualNet's separate slow self-supervised learner and fast adaptation learner | Continual learning with ResNet18 · CORE50-TA | ACC: 54.17±0.85 -> 57.64±1.36 | Improved |
| Use DualNet's separate slow self-supervised learner and fast adaptation learner | Continual learning with ResNet18 · CORE50-TF | ACC: 28.10±0.80 -> 38.76±1.52 | Improved |
| Use the slow learner on unlabeled stream samples | DualNet · Split miniImageNet-TA, ρ = 10% | ACC: 49.80±2.66 -> 54.03±2.88 | Improved |
| Use the slow learner on unlabeled stream samples | DualNet · Split miniImageNet-TA, ρ = 25% | ACC: 55.90±0.86 -> 62.80±2.40 | Improved |
| Replace Look-ahead with SGD for the Barlow Twins slow learner (ablation) | DualNet · Split miniImageNet-TA | ACC: 73.20±0.68 -> 64.20/-2.37 | Hurt |
| Use SimCLR instead of Barlow Twins with Look-ahead (ablation) | DualNet · Split miniImageNet-TA | ACC: 73.20±0.68 -> 72.13±0.44 | Hurt |
| Replace Barlow Twins with classification loss using Look-ahead (ablation) | DualNet · Split miniImageNet-TA | ACC: 73.20±0.68 -> 70.96±1.08 | Hurt |
| Use Look-ahead instead of SGD for the slow learner (ablation) | DualNet with Barlow Twins · Split miniImageNet-TA | ACC: 64.20/-2.37 -> 73.20±0.68 | Improved |
| Increase SSL updates between labeled batches from n=1 to n=20 (ablation) | DualNet, memory 50 · Split miniImageNet-TA | ACC: 72.26±0.71 -> 74.53±1.18 | Improved |
| Increase SSL updates between labeled batches from n=1 to n=20 (ablation) | DualNet, memory 100 · Split miniImageNet-TA | ACC: 74.67±1.05 -> 77.86±0.43 | Improved |
Later papers that compared with this method
2 later accepted papers report DualNet as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Expandable and Differentiable Dual Memories with Orthogonal Regularization for Exemplar-free Continual Learning (AAAI 2026 · better on 5 of 5 shared results)
- New Insights for the Stability-Plasticity Dilemma in Online Continual Learning (ICLR 2023 · better on 4 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Continual Learning, Fast and Slow (TPAMI 2023 · similarity 0.98)
- Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System (ICLR 2022 · similarity 0.89)
- CRNet: A Fast Continual Learning Framework With Random Theory (TPAMI 2023 · similarity 0.88)
- Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning (NeurIPS 2025 · similarity 0.86)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.86)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
